arch/ia64/sn/kernel/xpc_main.c

   1 /*
   2  * This file is subject to the terms and conditions of the GNU General Public
   3  * License.  See the file "COPYING" in the main directory of this archive
   4  * for more details.
   5  *
   6  * Copyright (c) 2004-2005 Silicon Graphics, Inc.  All Rights Reserved.
   7  */
   8
   9
  10 /*
  11  * Cross Partition Communication (XPC) support - standard version.
  12  *
  13  *      XPC provides a message passing capability that crosses partition
  14  *      boundaries. This module is made up of two parts:
  15  *
  16  *          partition   This part detects the presence/absence of other
  17  *                      partitions. It provides a heartbeat and monitors
  18  *                      the heartbeats of other partitions.
  19  *
  20  *          channel     This part manages the channels and sends/receives
  21  *                      messages across them to/from other partitions.
  22  *
  23  *      There are a couple of additional functions residing in XP, which
  24  *      provide an interface to XPC for its users.
  25  *
  26  *
  27  *      Caveats:
  28  *
  29  *        . We currently have no way to determine which nasid an IPI came
  30  *          from. Thus, xpc_IPI_send() does a remote AMO write followed by
  31  *          an IPI. The AMO indicates where data is to be pulled from, so
  32  *          after the IPI arrives, the remote partition checks the AMO word.
  33  *          The IPI can actually arrive before the AMO however, so other code
  34  *          must periodically check for this case. Also, remote AMO operations
  35  *          do not reliably time out. Thus we do a remote PIO read solely to
  36  *          know whether the remote partition is down and whether we should
  37  *          stop sending IPIs to it. This remote PIO read operation is set up
  38  *          in a special nofault region so SAL knows to ignore (and cleanup)
  39  *          any errors due to the remote AMO write, PIO read, and/or PIO
  40  *          write operations.
  41  *
  42  *          If/when new hardware solves this IPI problem, we should abandon
  43  *          the current approach.
  44  *
  45  */
  46
  47
  48 #include <linux/kernel.h>
  49 #include <linux/module.h>
  50 #include <linux/init.h>
  51 #include <linux/sched.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/cache.h>
  54 #include <linux/interrupt.h>
  55 #include <linux/slab.h>
  56 #include <linux/delay.h>
  57 #include <linux/reboot.h>
  58 #include <asm/sn/intr.h>
  59 #include <asm/sn/sn_sal.h>
  60 #include <asm/uaccess.h>
  61 #include "xpc.h"
  62
  63
  64 /* define two XPC debug device structures to be used with dev_dbg() et al */
  65
  66 struct device_driver xpc_dbg_name = {
  67         .name = "xpc"
  68 };
  69
  70 struct device xpc_part_dbg_subname = {
  71         .bus_id = {0},          /* set to "part" at xpc_init() time */
  72         .driver = &xpc_dbg_name
  73 };
  74
  75 struct device xpc_chan_dbg_subname = {
  76         .bus_id = {0},          /* set to "chan" at xpc_init() time */
  77         .driver = &xpc_dbg_name
  78 };
  79
  80 struct device *xpc_part = &xpc_part_dbg_subname;
  81 struct device *xpc_chan = &xpc_chan_dbg_subname;
  82
  83
  84 /* systune related variables for /proc/sys directories */
  85
  86 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
  87 static int xpc_hb_min_interval = 1;
  88 static int xpc_hb_max_interval = 10;
  89
  90 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
  91 static int xpc_hb_check_min_interval = 10;
  92 static int xpc_hb_check_max_interval = 120;
  93
  94 int xpc_disengage_request_timelimit = XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT;
  95 static int xpc_disengage_request_min_timelimit = 0;
  96 static int xpc_disengage_request_max_timelimit = 120;
  97
  98 static ctl_table xpc_sys_xpc_hb_dir[] = {
  99         {
 100                 1,
 101                 "hb_interval",
 102                 &xpc_hb_interval,
 103                 sizeof(int),
 104                 0644,
 105                 NULL,
 106                 &proc_dointvec_minmax,
 107                 &sysctl_intvec,
 108                 NULL,
 109                 &xpc_hb_min_interval,
 110                 &xpc_hb_max_interval
 111         },
 112         {
 113                 2,
 114                 "hb_check_interval",
 115                 &xpc_hb_check_interval,
 116                 sizeof(int),
 117                 0644,
 118                 NULL,
 119                 &proc_dointvec_minmax,
 120                 &sysctl_intvec,
 121                 NULL,
 122                 &xpc_hb_check_min_interval,
 123                 &xpc_hb_check_max_interval
 124         },
 125         {0}
 126 };
 127 static ctl_table xpc_sys_xpc_dir[] = {
 128         {
 129                 1,
 130                 "hb",
 131                 NULL,
 132                 0,
 133                 0555,
 134                 xpc_sys_xpc_hb_dir
 135         },
 136         {
 137                 2,
 138                 "disengage_request_timelimit",
 139                 &xpc_disengage_request_timelimit,
 140                 sizeof(int),
 141                 0644,
 142                 NULL,
 143                 &proc_dointvec_minmax,
 144                 &sysctl_intvec,
 145                 NULL,
 146                 &xpc_disengage_request_min_timelimit,
 147                 &xpc_disengage_request_max_timelimit
 148         },
 149         {0}
 150 };
 151 static ctl_table xpc_sys_dir[] = {
 152         {
 153                 1,
 154                 "xpc",
 155                 NULL,
 156                 0,
 157                 0555,
 158                 xpc_sys_xpc_dir
 159         },
 160         {0}
 161 };
 162 static struct ctl_table_header *xpc_sysctl;
 163
 164
 165 /* #of IRQs received */
 166 static atomic_t xpc_act_IRQ_rcvd;
 167
 168 /* IRQ handler notifies this wait queue on receipt of an IRQ */
 169 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq);
 170
 171 static unsigned long xpc_hb_check_timeout;
 172
 173 /* notification that the xpc_hb_checker thread has exited */
 174 static DECLARE_MUTEX_LOCKED(xpc_hb_checker_exited);
 175
 176 /* notification that the xpc_discovery thread has exited */
 177 static DECLARE_MUTEX_LOCKED(xpc_discovery_exited);
 178
 179
 180 static struct timer_list xpc_hb_timer;
 181
 182
 183 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
 184
 185
 186 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
 187 static struct notifier_block xpc_reboot_notifier = {
 188         .notifier_call = xpc_system_reboot,
 189 };
 190
 191
 192 /*
 193  * Timer function to enforce the timelimit on the partition disengage request.
 194  */
 195 static void
 196 xpc_timeout_partition_disengage_request(unsigned long data)
 197 {
 198         struct xpc_partition *part = (struct xpc_partition *) data;
 199
 200
 201         DBUG_ON(jiffies < part->disengage_request_timeout);
 202
 203         (void) xpc_partition_disengaged(part);
 204
 205         DBUG_ON(part->disengage_request_timeout != 0);
 206         DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part)) != 0);
 207 }
 208
 209
 210 /*
 211  * Notify the heartbeat check thread that an IRQ has been received.
 212  */
 213 static irqreturn_t
 214 xpc_act_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
 215 {
 216         atomic_inc(&xpc_act_IRQ_rcvd);
 217         wake_up_interruptible(&xpc_act_IRQ_wq);
 218         return IRQ_HANDLED;
 219 }
 220
 221
 222 /*
 223  * Timer to produce the heartbeat.  The timer structures function is
 224  * already set when this is initially called.  A tunable is used to
 225  * specify when the next timeout should occur.
 226  */
 227 static void
 228 xpc_hb_beater(unsigned long dummy)
 229 {
 230         xpc_vars->heartbeat++;
 231
 232         if (jiffies >= xpc_hb_check_timeout) {
 233                 wake_up_interruptible(&xpc_act_IRQ_wq);
 234         }
 235
 236         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
 237         add_timer(&xpc_hb_timer);
 238 }
 239
 240
 241 /*
 242  * This thread is responsible for nearly all of the partition
 243  * activation/deactivation.
 244  */
 245 static int
 246 xpc_hb_checker(void *ignore)
 247 {
 248         int last_IRQ_count = 0;
 249         int new_IRQ_count;
 250         int force_IRQ=0;
 251
 252
 253         /* this thread was marked active by xpc_hb_init() */
 254
 255         daemonize(XPC_HB_CHECK_THREAD_NAME);
 256
 257         set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU));
 258
 259         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
 260
 261         while (!(volatile int) xpc_exiting) {
 262
 263                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
 264                         "been received\n",
 265                         (int) (xpc_hb_check_timeout - jiffies),
 266                         atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count);
 267
 268
 269                 /* checking of remote heartbeats is skewed by IRQ handling */
 270                 if (jiffies >= xpc_hb_check_timeout) {
 271                         dev_dbg(xpc_part, "checking remote heartbeats\n");
 272                         xpc_check_remote_hb();
 273
 274                         /*
 275                          * We need to periodically recheck to ensure no
 276                          * IPI/AMO pairs have been missed.  That check
 277                          * must always reset xpc_hb_check_timeout.
 278                          */
 279                         force_IRQ = 1;
 280                 }
 281
 282
 283                 /* check for outstanding IRQs */
 284                 new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd);
 285                 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
 286                         force_IRQ = 0;
 287
 288                         dev_dbg(xpc_part, "found an IRQ to process; will be "
 289                                 "resetting xpc_hb_check_timeout\n");
 290
 291                         last_IRQ_count += xpc_identify_act_IRQ_sender();
 292                         if (last_IRQ_count < new_IRQ_count) {
 293                                 /* retry once to help avoid missing AMO */
 294                                 (void) xpc_identify_act_IRQ_sender();
 295                         }
 296                         last_IRQ_count = new_IRQ_count;
 297
 298                         xpc_hb_check_timeout = jiffies +
 299                                            (xpc_hb_check_interval * HZ);
 300                 }
 301
 302                 /* wait for IRQ or timeout */
 303                 (void) wait_event_interruptible(xpc_act_IRQ_wq,
 304                             (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) ||
 305                                         jiffies >= xpc_hb_check_timeout ||
 306                                                 (volatile int) xpc_exiting));
 307         }
 308
 309         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
 310
 311
 312         /* mark this thread as having exited */
 313         up(&xpc_hb_checker_exited);
 314         return 0;
 315 }
 316
 317
 318 /*
 319  * This thread will attempt to discover other partitions to activate
 320  * based on info provided by SAL. This new thread is short lived and
 321  * will exit once discovery is complete.
 322  */
 323 static int
 324 xpc_initiate_discovery(void *ignore)
 325 {
 326         daemonize(XPC_DISCOVERY_THREAD_NAME);
 327
 328         xpc_discovery();
 329
 330         dev_dbg(xpc_part, "discovery thread is exiting\n");
 331
 332         /* mark this thread as having exited */
 333         up(&xpc_discovery_exited);
 334         return 0;
 335 }
 336
 337
 338 /*
 339  * Establish first contact with the remote partititon. This involves pulling
 340  * the XPC per partition variables from the remote partition and waiting for
 341  * the remote partition to pull ours.
 342  */
 343 static enum xpc_retval
 344 xpc_make_first_contact(struct xpc_partition *part)
 345 {
 346         enum xpc_retval ret;
 347
 348
 349         while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) {
 350                 if (ret != xpcRetry) {
 351                         XPC_DEACTIVATE_PARTITION(part, ret);
 352                         return ret;
 353                 }
 354
 355                 dev_dbg(xpc_chan, "waiting to make first contact with "
 356                         "partition %d\n", XPC_PARTID(part));
 357
 358                 /* wait a 1/4 of a second or so */
 359                 (void) msleep_interruptible(250);
 360
 361                 if (part->act_state == XPC_P_DEACTIVATING) {
 362                         return part->reason;
 363                 }
 364         }
 365
 366         return xpc_mark_partition_active(part);
 367 }
 368
 369
 370 /*
 371  * The first kthread assigned to a newly activated partition is the one
 372  * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
 373  * that kthread until the partition is brought down, at which time that kthread
 374  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
 375  * that XPC has dismantled all communication infrastructure for the associated
 376  * partition.) This kthread becomes the channel manager for that partition.
 377  *
 378  * Each active partition has a channel manager, who, besides connecting and
 379  * disconnecting channels, will ensure that each of the partition's connected
 380  * channels has the required number of assigned kthreads to get the work done.
 381  */
 382 static void
 383 xpc_channel_mgr(struct xpc_partition *part)
 384 {
 385         while (part->act_state != XPC_P_DEACTIVATING ||
 386                         atomic_read(&part->nchannels_active) > 0 ||
 387                                         !xpc_partition_disengaged(part)) {
 388
 389                 xpc_process_channel_activity(part);
 390
 391
 392                 /*
 393                  * Wait until we've been requested to activate kthreads or
 394                  * all of the channel's message queues have been torn down or
 395                  * a signal is pending.
 396                  *
 397                  * The channel_mgr_requests is set to 1 after being awakened,
 398                  * This is done to prevent the channel mgr from making one pass
 399                  * through the loop for each request, since he will
 400                  * be servicing all the requests in one pass. The reason it's
 401                  * set to 1 instead of 0 is so that other kthreads will know
 402                  * that the channel mgr is running and won't bother trying to
 403                  * wake him up.
 404                  */
 405                 atomic_dec(&part->channel_mgr_requests);
 406                 (void) wait_event_interruptible(part->channel_mgr_wq,
 407                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
 408                                 (volatile u64) part->local_IPI_amo != 0 ||
 409                                 ((volatile u8) part->act_state ==
 410                                                         XPC_P_DEACTIVATING &&
 411                                 atomic_read(&part->nchannels_active) == 0 &&
 412                                 xpc_partition_disengaged(part))));
 413                 atomic_set(&part->channel_mgr_requests, 1);
 414
 415                 // >>> Does it need to wakeup periodically as well? In case we
 416                 // >>> miscalculated the #of kthreads to wakeup or create?
 417         }
 418 }
 419
 420
 421 /*
 422  * When XPC HB determines that a partition has come up, it will create a new
 423  * kthread and that kthread will call this function to attempt to set up the
 424  * basic infrastructure used for Cross Partition Communication with the newly
 425  * upped partition.
 426  *
 427  * The kthread that was created by XPC HB and which setup the XPC
 428  * infrastructure will remain assigned to the partition until the partition
 429  * goes down. At which time the kthread will teardown the XPC infrastructure
 430  * and then exit.
 431  *
 432  * XPC HB will put the remote partition's XPC per partition specific variables
 433  * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
 434  * calling xpc_partition_up().
 435  */
 436 static void
 437 xpc_partition_up(struct xpc_partition *part)
 438 {
 439         DBUG_ON(part->channels != NULL);
 440
 441         dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part));
 442
 443         if (xpc_setup_infrastructure(part) != xpcSuccess) {
 444                 return;
 445         }
 446
 447         /*
 448          * The kthread that XPC HB called us with will become the
 449          * channel manager for this partition. It will not return
 450          * back to XPC HB until the partition's XPC infrastructure
 451          * has been dismantled.
 452          */
 453
 454         (void) xpc_part_ref(part);      /* this will always succeed */
 455
 456         if (xpc_make_first_contact(part) == xpcSuccess) {
 457                 xpc_channel_mgr(part);
 458         }
 459
 460         xpc_part_deref(part);
 461
 462         xpc_teardown_infrastructure(part);
 463 }
 464
 465
 466 static int
 467 xpc_activating(void *__partid)
 468 {
 469         partid_t partid = (u64) __partid;
 470         struct xpc_partition *part = &xpc_partitions[partid];
 471         unsigned long irq_flags;
 472         struct sched_param param = { sched_priority: MAX_RT_PRIO - 1 };
 473         int ret;
 474
 475
 476         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
 477
 478         spin_lock_irqsave(&part->act_lock, irq_flags);
 479
 480         if (part->act_state == XPC_P_DEACTIVATING) {
 481                 part->act_state = XPC_P_INACTIVE;
 482                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 483                 part->remote_rp_pa = 0;
 484                 return 0;
 485         }
 486
 487         /* indicate the thread is activating */
 488         DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
 489         part->act_state = XPC_P_ACTIVATING;
 490
 491         XPC_SET_REASON(part, 0, 0);
 492         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 493
 494         dev_dbg(xpc_part, "bringing partition %d up\n", partid);
 495
 496         daemonize("xpc%02d", partid);
 497
 498         /*
 499          * This thread needs to run at a realtime priority to prevent a
 500          * significant performance degradation.
 501          */
 502         ret = sched_setscheduler(current, SCHED_FIFO, &param);
 503         if (ret != 0) {
 504                 dev_warn(xpc_part, "unable to set pid %d to a realtime "
 505                         "priority, ret=%d\n", current->pid, ret);
 506         }
 507
 508         /* allow this thread and its children to run on any CPU */
 509         set_cpus_allowed(current, CPU_MASK_ALL);
 510
 511         /*
 512          * Register the remote partition's AMOs with SAL so it can handle
 513          * and cleanup errors within that address range should the remote
 514          * partition go down. We don't unregister this range because it is
 515          * difficult to tell when outstanding writes to the remote partition
 516          * are finished and thus when it is safe to unregister. This should
 517          * not result in wasted space in the SAL xp_addr_region table because
 518          * we should get the same page for remote_amos_page_pa after module
 519          * reloads and system reboots.
 520          */
 521         if (sn_register_xp_addr_region(part->remote_amos_page_pa,
 522                                                         PAGE_SIZE, 1) < 0) {
 523                 dev_warn(xpc_part, "xpc_partition_up(%d) failed to register "
 524                         "xp_addr region\n", partid);
 525
 526                 spin_lock_irqsave(&part->act_lock, irq_flags);
 527                 part->act_state = XPC_P_INACTIVE;
 528                 XPC_SET_REASON(part, xpcPhysAddrRegFailed, __LINE__);
 529                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 530                 part->remote_rp_pa = 0;
 531                 return 0;
 532         }
 533
 534         xpc_allow_hb(partid, xpc_vars);
 535         xpc_IPI_send_activated(part);
 536
 537
 538         /*
 539          * xpc_partition_up() holds this thread and marks this partition as
 540          * XPC_P_ACTIVE by calling xpc_hb_mark_active().
 541          */
 542         (void) xpc_partition_up(part);
 543
 544         xpc_disallow_hb(partid, xpc_vars);
 545         xpc_mark_partition_inactive(part);
 546
 547         if (part->reason == xpcReactivating) {
 548                 /* interrupting ourselves results in activating partition */
 549                 xpc_IPI_send_reactivate(part);
 550         }
 551
 552         return 0;
 553 }
 554
 555
 556 void
 557 xpc_activate_partition(struct xpc_partition *part)
 558 {
 559         partid_t partid = XPC_PARTID(part);
 560         unsigned long irq_flags;
 561         pid_t pid;
 562
 563
 564         spin_lock_irqsave(&part->act_lock, irq_flags);
 565
 566         pid = kernel_thread(xpc_activating, (void *) ((u64) partid), 0);
 567
 568         DBUG_ON(part->act_state != XPC_P_INACTIVE);
 569
 570         if (pid > 0) {
 571                 part->act_state = XPC_P_ACTIVATION_REQ;
 572                 XPC_SET_REASON(part, xpcCloneKThread, __LINE__);
 573         } else {
 574                 XPC_SET_REASON(part, xpcCloneKThreadFailed, __LINE__);
 575         }
 576
 577         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 578 }
 579
 580
 581 /*
 582  * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
 583  * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
 584  * than one partition, we use an AMO_t structure per partition to indicate
 585  * whether a partition has sent an IPI or not.  >>> If it has, then wake up the
 586  * associated kthread to handle it.
 587  *
 588  * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
 589  * running on other partitions.
 590  *
 591  * Noteworthy Arguments:
 592  *
 593  *      irq - Interrupt ReQuest number. NOT USED.
 594  *
 595  *      dev_id - partid of IPI's potential sender.
 596  *
 597  *      regs - processor's context before the processor entered
 598  *             interrupt code. NOT USED.
 599  */
 600 irqreturn_t
 601 xpc_notify_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
 602 {
 603         partid_t partid = (partid_t) (u64) dev_id;
 604         struct xpc_partition *part = &xpc_partitions[partid];
 605
 606
 607         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
 608
 609         if (xpc_part_ref(part)) {
 610                 xpc_check_for_channel_activity(part);
 611
 612                 xpc_part_deref(part);
 613         }
 614         return IRQ_HANDLED;
 615 }
 616
 617
 618 /*
 619  * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
 620  * because the write to their associated IPI amo completed after the IRQ/IPI
 621  * was received.
 622  */
 623 void
 624 xpc_dropped_IPI_check(struct xpc_partition *part)
 625 {
 626         if (xpc_part_ref(part)) {
 627                 xpc_check_for_channel_activity(part);
 628
 629                 part->dropped_IPI_timer.expires = jiffies +
 630                                                         XPC_P_DROPPED_IPI_WAIT;
 631                 add_timer(&part->dropped_IPI_timer);
 632                 xpc_part_deref(part);
 633         }
 634 }
 635
 636
 637 void
 638 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
 639 {
 640         int idle = atomic_read(&ch->kthreads_idle);
 641         int assigned = atomic_read(&ch->kthreads_assigned);
 642         int wakeup;
 643
 644
 645         DBUG_ON(needed <= 0);
 646
 647         if (idle > 0) {
 648                 wakeup = (needed > idle) ? idle : needed;
 649                 needed -= wakeup;
 650
 651                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
 652                         "channel=%d\n", wakeup, ch->partid, ch->number);
 653
 654                 /* only wakeup the requested number of kthreads */
 655                 wake_up_nr(&ch->idle_wq, wakeup);
 656         }
 657
 658         if (needed <= 0) {
 659                 return;
 660         }
 661
 662         if (needed + assigned > ch->kthreads_assigned_limit) {
 663                 needed = ch->kthreads_assigned_limit - assigned;
 664                 // >>>should never be less than 0
 665                 if (needed <= 0) {
 666                         return;
 667                 }
 668         }
 669
 670         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
 671                 needed, ch->partid, ch->number);
 672
 673         xpc_create_kthreads(ch, needed);
 674 }
 675
 676
 677 /*
 678  * This function is where XPC's kthreads wait for messages to deliver.
 679  */
 680 static void
 681 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
 682 {
 683         do {
 684                 /* deliver messages to their intended recipients */
 685
 686                 while ((volatile s64) ch->w_local_GP.get <
 687                                 (volatile s64) ch->w_remote_GP.put &&
 688                                         !((volatile u32) ch->flags &
 689                                                 XPC_C_DISCONNECTING)) {
 690                         xpc_deliver_msg(ch);
 691                 }
 692
 693                 if (atomic_inc_return(&ch->kthreads_idle) >
 694                                                 ch->kthreads_idle_limit) {
 695                         /* too many idle kthreads on this channel */
 696                         atomic_dec(&ch->kthreads_idle);
 697                         break;
 698                 }
 699
 700                 dev_dbg(xpc_chan, "idle kthread calling "
 701                         "wait_event_interruptible_exclusive()\n");
 702
 703                 (void) wait_event_interruptible_exclusive(ch->idle_wq,
 704                                 ((volatile s64) ch->w_local_GP.get <
 705                                         (volatile s64) ch->w_remote_GP.put ||
 706                                 ((volatile u32) ch->flags &
 707                                                 XPC_C_DISCONNECTING)));
 708
 709                 atomic_dec(&ch->kthreads_idle);
 710
 711         } while (!((volatile u32) ch->flags & XPC_C_DISCONNECTING));
 712 }
 713
 714
 715 static int
 716 xpc_daemonize_kthread(void *args)
 717 {
 718         partid_t partid = XPC_UNPACK_ARG1(args);
 719         u16 ch_number = XPC_UNPACK_ARG2(args);
 720         struct xpc_partition *part = &xpc_partitions[partid];
 721         struct xpc_channel *ch;
 722         int n_needed;
 723         unsigned long irq_flags;
 724
 725
 726         daemonize("xpc%02dc%d", partid, ch_number);
 727
 728         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
 729                 partid, ch_number);
 730
 731         ch = &part->channels[ch_number];
 732
 733         if (!(ch->flags & XPC_C_DISCONNECTING)) {
 734
 735                 /* let registerer know that connection has been established */
 736
 737                 spin_lock_irqsave(&ch->lock, irq_flags);
 738                 if (!(ch->flags & XPC_C_CONNECTCALLOUT)) {
 739                         ch->flags |= XPC_C_CONNECTCALLOUT;
 740                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 741
 742                         xpc_connected_callout(ch);
 743
 744                         /*
 745                          * It is possible that while the callout was being
 746                          * made that the remote partition sent some messages.
 747                          * If that is the case, we may need to activate
 748                          * additional kthreads to help deliver them. We only
 749                          * need one less than total #of messages to deliver.
 750                          */
 751                         n_needed = ch->w_remote_GP.put - ch->w_local_GP.get - 1;
 752                         if (n_needed > 0 &&
 753                                         !(ch->flags & XPC_C_DISCONNECTING)) {
 754                                 xpc_activate_kthreads(ch, n_needed);
 755                         }
 756                 } else {
 757                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 758                 }
 759
 760                 xpc_kthread_waitmsgs(part, ch);
 761         }
 762
 763         if (atomic_dec_return(&ch->kthreads_assigned) == 0) {
 764                 spin_lock_irqsave(&ch->lock, irq_flags);
 765                 if ((ch->flags & XPC_C_CONNECTCALLOUT) &&
 766                                 !(ch->flags & XPC_C_DISCONNECTCALLOUT)) {
 767                         ch->flags |= XPC_C_DISCONNECTCALLOUT;
 768                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 769
 770                         xpc_disconnecting_callout(ch);
 771                 } else {
 772                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 773                 }
 774                 if (atomic_dec_return(&part->nchannels_engaged) == 0) {
 775                         xpc_mark_partition_disengaged(part);
 776                         xpc_IPI_send_disengage(part);
 777                 }
 778         }
 779
 780
 781         xpc_msgqueue_deref(ch);
 782
 783         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
 784                 partid, ch_number);
 785
 786         xpc_part_deref(part);
 787         return 0;
 788 }
 789
 790
 791 /*
 792  * For each partition that XPC has established communications with, there is
 793  * a minimum of one kernel thread assigned to perform any operation that
 794  * may potentially sleep or block (basically the callouts to the asynchronous
 795  * functions registered via xpc_connect()).
 796  *
 797  * Additional kthreads are created and destroyed by XPC as the workload
 798  * demands.
 799  *
 800  * A kthread is assigned to one of the active channels that exists for a given
 801  * partition.
 802  */
 803 void
 804 xpc_create_kthreads(struct xpc_channel *ch, int needed)
 805 {
 806         unsigned long irq_flags;
 807         pid_t pid;
 808         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
 809         struct xpc_partition *part = &xpc_partitions[ch->partid];
 810
 811
 812         while (needed-- > 0) {
 813
 814                 /*
 815                  * The following is done on behalf of the newly created
 816                  * kthread. That kthread is responsible for doing the
 817                  * counterpart to the following before it exits.
 818                  */
 819                 (void) xpc_part_ref(part);
 820                 xpc_msgqueue_ref(ch);
 821                 if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
 822                     atomic_inc_return(&part->nchannels_engaged) == 1) {
 823                         xpc_mark_partition_engaged(part);
 824                 }
 825
 826                 pid = kernel_thread(xpc_daemonize_kthread, (void *) args, 0);
 827                 if (pid < 0) {
 828                         /* the fork failed */
 829                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 830                             atomic_dec_return(&part->nchannels_engaged) == 0) {
 831                                 xpc_mark_partition_disengaged(part);
 832                                 xpc_IPI_send_disengage(part);
 833                         }
 834                         xpc_msgqueue_deref(ch);
 835                         xpc_part_deref(part);
 836
 837                         if (atomic_read(&ch->kthreads_assigned) <
 838                                                 ch->kthreads_idle_limit) {
 839                                 /*
 840                                  * Flag this as an error only if we have an
 841                                  * insufficient #of kthreads for the channel
 842                                  * to function.
 843                                  *
 844                                  * No xpc_msgqueue_ref() is needed here since
 845                                  * the channel mgr is doing this.
 846                                  */
 847                                 spin_lock_irqsave(&ch->lock, irq_flags);
 848                                 XPC_DISCONNECT_CHANNEL(ch, xpcLackOfResources,
 849                                                                 &irq_flags);
 850                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 851                         }
 852                         break;
 853                 }
 854
 855                 ch->kthreads_created++; // >>> temporary debug only!!!
 856         }
 857 }
 858
 859
 860 void
 861 xpc_disconnect_wait(int ch_number)
 862 {
 863         unsigned long irq_flags;
 864         partid_t partid;
 865         struct xpc_partition *part;
 866         struct xpc_channel *ch;
 867         int wakeup_channel_mgr;
 868
 869
 870         /* now wait for all callouts to the caller's function to cease */
 871         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
 872                 part = &xpc_partitions[partid];
 873
 874                 if (!xpc_part_ref(part)) {
 875                         continue;
 876                 }
 877
 878                 ch = &part->channels[ch_number];
 879
 880                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
 881                         xpc_part_deref(part);
 882                         continue;
 883                 }
 884
 885                 (void) down(&ch->wdisconnect_sema);
 886
 887                 spin_lock_irqsave(&ch->lock, irq_flags);
 888                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
 889                 wakeup_channel_mgr = 0;
 890
 891                 if (ch->delayed_IPI_flags) {
 892                         if (part->act_state != XPC_P_DEACTIVATING) {
 893                                 spin_lock(&part->IPI_lock);
 894                                 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
 895                                         ch->number, ch->delayed_IPI_flags);
 896                                 spin_unlock(&part->IPI_lock);
 897                                 wakeup_channel_mgr = 1;
 898                         }
 899                         ch->delayed_IPI_flags = 0;
 900                 }
 901
 902                 ch->flags &= ~XPC_C_WDISCONNECT;
 903                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 904
 905                 if (wakeup_channel_mgr) {
 906                         xpc_wakeup_channel_mgr(part);
 907                 }
 908
 909                 xpc_part_deref(part);
 910         }
 911 }
 912
 913
 914 static void
 915 xpc_do_exit(enum xpc_retval reason)
 916 {
 917         partid_t partid;
 918         int active_part_count;
 919         struct xpc_partition *part;
 920         unsigned long printmsg_time;
 921
 922
 923         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
 924         DBUG_ON(xpc_exiting == 1);
 925
 926         /*
 927          * Let the heartbeat checker thread and the discovery thread
 928          * (if one is running) know that they should exit. Also wake up
 929          * the heartbeat checker thread in case it's sleeping.
 930          */
 931         xpc_exiting = 1;
 932         wake_up_interruptible(&xpc_act_IRQ_wq);
 933
 934         /* ignore all incoming interrupts */
 935         free_irq(SGI_XPC_ACTIVATE, NULL);
 936
 937         /* wait for the discovery thread to exit */
 938         down(&xpc_discovery_exited);
 939
 940         /* wait for the heartbeat checker thread to exit */
 941         down(&xpc_hb_checker_exited);
 942
 943
 944         /* sleep for a 1/3 of a second or so */
 945         (void) msleep_interruptible(300);
 946
 947
 948         /* wait for all partitions to become inactive */
 949
 950         printmsg_time = jiffies;
 951
 952         do {
 953                 active_part_count = 0;
 954
 955                 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
 956                         part = &xpc_partitions[partid];
 957
 958                         if (xpc_partition_disengaged(part) &&
 959                                         part->act_state == XPC_P_INACTIVE) {
 960                                 continue;
 961                         }
 962
 963                         active_part_count++;
 964
 965                         XPC_DEACTIVATE_PARTITION(part, reason);
 966                 }
 967
 968                 if (active_part_count == 0) {
 969                         break;
 970                 }
 971
 972                 if (jiffies >= printmsg_time) {
 973                         dev_info(xpc_part, "waiting for partitions to "
 974                                 "deactivate/disengage, active count=%d, remote "
 975                                 "engaged=0x%lx\n", active_part_count,
 976                                 xpc_partition_engaged(1UL << partid));
 977
 978                         printmsg_time = jiffies +
 979                                         (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
 980                 }
 981
 982                 /* sleep for a 1/3 of a second or so */
 983                 (void) msleep_interruptible(300);
 984
 985         } while (1);
 986
 987         DBUG_ON(xpc_partition_engaged(-1UL));
 988
 989
 990         /* indicate to others that our reserved page is uninitialized */
 991         xpc_rsvd_page->vars_pa = 0;
 992
 993         /* now it's time to eliminate our heartbeat */
 994         del_timer_sync(&xpc_hb_timer);
 995         DBUG_ON(xpc_vars->heartbeating_to_mask != 0);
 996
 997         /* take ourselves off of the reboot_notifier_list */
 998         (void) unregister_reboot_notifier(&xpc_reboot_notifier);
 999
1000         /* close down protections for IPI operations */
1001         xpc_restrict_IPI_ops();
1002
1003
1004         /* clear the interface to XPC's functions */
1005         xpc_clear_interface();
1006
1007         if (xpc_sysctl) {
1008                 unregister_sysctl_table(xpc_sysctl);
1009         }
1010 }
1011
1012
1013 /*
1014  * This function is called when the system is being rebooted.
1015  */
1016 static int
1017 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1018 {
1019         enum xpc_retval reason;
1020
1021
1022         switch (event) {
1023         case SYS_RESTART:
1024                 reason = xpcSystemReboot;
1025                 break;
1026         case SYS_HALT:
1027                 reason = xpcSystemHalt;
1028                 break;
1029         case SYS_POWER_OFF:
1030                 reason = xpcSystemPoweroff;
1031                 break;
1032         default:
1033                 reason = xpcSystemGoingDown;
1034         }
1035
1036         xpc_do_exit(reason);
1037         return NOTIFY_DONE;
1038 }
1039
1040
1041 int __init
1042 xpc_init(void)
1043 {
1044         int ret;
1045         partid_t partid;
1046         struct xpc_partition *part;
1047         pid_t pid;
1048
1049
1050         if (!ia64_platform_is("sn2")) {
1051                 return -ENODEV;
1052         }
1053
1054         /*
1055          * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng
1056          * various portions of a partition's reserved page. Its size is based
1057          * on the size of the reserved page header and part_nasids mask. So we
1058          * need to ensure that the other items will fit as well.
1059          */
1060         if (XPC_RP_VARS_SIZE > XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES) {
1061                 dev_err(xpc_part, "xpc_remote_copy_buffer is not big enough\n");
1062                 return -EPERM;
1063         }
1064         DBUG_ON((u64) xpc_remote_copy_buffer !=
1065                                 L1_CACHE_ALIGN((u64) xpc_remote_copy_buffer));
1066
1067         snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1068         snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1069
1070         xpc_sysctl = register_sysctl_table(xpc_sys_dir, 1);
1071
1072         /*
1073          * The first few fields of each entry of xpc_partitions[] need to
1074          * be initialized now so that calls to xpc_connect() and
1075          * xpc_disconnect() can be made prior to the activation of any remote
1076          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1077          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1078          * PARTITION HAS BEEN ACTIVATED.
1079          */
1080         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1081                 part = &xpc_partitions[partid];
1082
1083                 DBUG_ON((u64) part != L1_CACHE_ALIGN((u64) part));
1084
1085                 part->act_IRQ_rcvd = 0;
1086                 spin_lock_init(&part->act_lock);
1087                 part->act_state = XPC_P_INACTIVE;
1088                 XPC_SET_REASON(part, 0, 0);
1089
1090                 init_timer(&part->disengage_request_timer);
1091                 part->disengage_request_timer.function =
1092                                 xpc_timeout_partition_disengage_request;
1093                 part->disengage_request_timer.data = (unsigned long) part;
1094
1095                 part->setup_state = XPC_P_UNSET;
1096                 init_waitqueue_head(&part->teardown_wq);
1097                 atomic_set(&part->references, 0);
1098         }
1099
1100         /*
1101          * Open up protections for IPI operations (and AMO operations on
1102          * Shub 1.1 systems).
1103          */
1104         xpc_allow_IPI_ops();
1105
1106         /*
1107          * Interrupts being processed will increment this atomic variable and
1108          * awaken the heartbeat thread which will process the interrupts.
1109          */
1110         atomic_set(&xpc_act_IRQ_rcvd, 0);
1111
1112         /*
1113          * This is safe to do before the xpc_hb_checker thread has started
1114          * because the handler releases a wait queue.  If an interrupt is
1115          * received before the thread is waiting, it will not go to sleep,
1116          * but rather immediately process the interrupt.
1117          */
1118         ret = request_irq(SGI_XPC_ACTIVATE, xpc_act_IRQ_handler, 0,
1119                                                         "xpc hb", NULL);
1120         if (ret != 0) {
1121                 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
1122                         "errno=%d\n", -ret);
1123
1124                 xpc_restrict_IPI_ops();
1125
1126                 if (xpc_sysctl) {
1127                         unregister_sysctl_table(xpc_sysctl);
1128                 }
1129                 return -EBUSY;
1130         }
1131
1132         /*
1133          * Fill the partition reserved page with the information needed by
1134          * other partitions to discover we are alive and establish initial
1135          * communications.
1136          */
1137         xpc_rsvd_page = xpc_rsvd_page_init();
1138         if (xpc_rsvd_page == NULL) {
1139                 dev_err(xpc_part, "could not setup our reserved page\n");
1140
1141                 free_irq(SGI_XPC_ACTIVATE, NULL);
1142                 xpc_restrict_IPI_ops();
1143
1144                 if (xpc_sysctl) {
1145                         unregister_sysctl_table(xpc_sysctl);
1146                 }
1147                 return -EBUSY;
1148         }
1149
1150
1151         /* add ourselves to the reboot_notifier_list */
1152         ret = register_reboot_notifier(&xpc_reboot_notifier);
1153         if (ret != 0) {
1154                 dev_warn(xpc_part, "can't register reboot notifier\n");
1155         }
1156
1157
1158         /*
1159          * Set the beating to other partitions into motion.  This is
1160          * the last requirement for other partitions' discovery to
1161          * initiate communications with us.
1162          */
1163         init_timer(&xpc_hb_timer);
1164         xpc_hb_timer.function = xpc_hb_beater;
1165         xpc_hb_beater(0);
1166
1167
1168         /*
1169          * The real work-horse behind xpc.  This processes incoming
1170          * interrupts and monitors remote heartbeats.
1171          */
1172         pid = kernel_thread(xpc_hb_checker, NULL, 0);
1173         if (pid < 0) {
1174                 dev_err(xpc_part, "failed while forking hb check thread\n");
1175
1176                 /* indicate to others that our reserved page is uninitialized */
1177                 xpc_rsvd_page->vars_pa = 0;
1178
1179                 /* take ourselves off of the reboot_notifier_list */
1180                 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1181
1182                 del_timer_sync(&xpc_hb_timer);
1183                 free_irq(SGI_XPC_ACTIVATE, NULL);
1184                 xpc_restrict_IPI_ops();
1185
1186                 if (xpc_sysctl) {
1187                         unregister_sysctl_table(xpc_sysctl);
1188                 }
1189                 return -EBUSY;
1190         }
1191
1192
1193         /*
1194          * Startup a thread that will attempt to discover other partitions to
1195          * activate based on info provided by SAL. This new thread is short
1196          * lived and will exit once discovery is complete.
1197          */
1198         pid = kernel_thread(xpc_initiate_discovery, NULL, 0);
1199         if (pid < 0) {
1200                 dev_err(xpc_part, "failed while forking discovery thread\n");
1201
1202                 /* mark this new thread as a non-starter */
1203                 up(&xpc_discovery_exited);
1204
1205                 xpc_do_exit(xpcUnloading);
1206                 return -EBUSY;
1207         }
1208
1209
1210         /* set the interface to point at XPC's functions */
1211         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1212                           xpc_initiate_allocate, xpc_initiate_send,
1213                           xpc_initiate_send_notify, xpc_initiate_received,
1214                           xpc_initiate_partid_to_nasids);
1215
1216         return 0;
1217 }
1218 module_init(xpc_init);
1219
1220
1221 void __exit
1222 xpc_exit(void)
1223 {
1224         xpc_do_exit(xpcUnloading);
1225 }
1226 module_exit(xpc_exit);
1227
1228
1229 MODULE_AUTHOR("Silicon Graphics, Inc.");
1230 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1231 MODULE_LICENSE("GPL");
1232
1233 module_param(xpc_hb_interval, int, 0);
1234 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1235                 "heartbeat increments.");
1236
1237 module_param(xpc_hb_check_interval, int, 0);
1238 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1239                 "heartbeat checks.");
1240
1241 module_param(xpc_disengage_request_timelimit, int, 0);
1242 MODULE_PARM_DESC(xpc_disengage_request_timelimit, "Number of seconds to wait "
1243                 "for disengage request to complete.");
1244